40 Trillion Tokens a Day (Yes, More Than OpenAI) | Lin Qiao, CEO of Fireworks
Lin Qiao, CEO of Fireworks, discusses why specialized AI models built on private company data are the future, contrasting them with general-purpose models. She argues for "open intelligence," revealing Fireworks processes over 40 trillion tokens daily from customized models, more than OpenAI's API. Qiao emphasizes the economic and strategic imperative for companies to own their specialized intelligence, advocating for open-sourcing by frontier labs like OpenAI and Anthropic, while detailing Fireworks' proprietary, quality-obsessed platform for tailored AI solutions.
Lin Qiao, CEO of Fireworks, argues against concerns about Chinese open-source AI models, advocating instead for an industry-wide push towards “open intelligence” and “open development.” She believes the AI industry is at a turning point, moving towards a future dominated not by a few general-purpose models, but by “millions of specialized models, one per application per use case.”
The Rise of Specialized Intelligence
Qiao’s core philosophy posits that intelligence is a derivative of data. While frontier labs train general-purpose AGI on public internet data, the vast majority of valuable data remains locked within enterprise applications. This private data, which constitutes a company’s “alpha,” should never be shared externally. Instead, it should be used to create customized models that are uniquely owned by the company.
Fireworks operates on the belief that this specialization process should be continuous—occurring weekly, daily, or even hourly—to keep pace with evolving applications and the rapid advancements in base models. This approach empowers companies with control over their IP and costs.
Fireworks’ Scale and Impact
Fireworks’ platform exemplifies this vision, processing over 40 trillion tokens a day, a volume Qiao suggests is greater than OpenAI’s and Gemini’s APIs combined. Notably, 95% of this traffic comes from customized models and inference deployments, underscoring the demand for specialized AI solutions.
Why Frontier Labs Don’t Prioritize Fine-Tuning
Qiao explains that for general intelligence companies, the unit economics don’t align with supporting extensive fine-tuning services. Training models from scratch is immensely expensive, and their business strategy focuses on rapidly scaling a few general models via API. Supporting millions of customized models would require a “completely different business” model and infrastructure.
The Importance of Reinforcement Learning for Customization
Fireworks offers varied levels of engagement for model customization, from deep, knob-level control for research teams (e.g., Cursor AI using RL on Fireworks’ training stack) to an SDK for developers with AI experience (e.g., Doximity in healthcare).
Reinforcement Learning (RL), in contrast to Supervised Fine-Tuning (SFT), allows a model to learn by interacting with a product or simulation, receiving rewards for good outcomes. This requires defining complex reward functions based on deep product knowledge and domain-specific judgment, which Qiao notes, is driving a new type of role blending product and research expertise.
The Strategic Value of Owning Intelligence
Drawing on a quote from Jensen Huang—“there’s no specialized general company”—Qiao emphasizes that every company is special, solving unique problems with unique solutions. Therefore, every company should own their intelligence, codifying their proprietary taste and judgment into their own models.
This ownership creates a strategic moat that cannot be easily replicated. Proprietary data reflecting customer intent, preferences, and business logic, when integrated into a company’s model, forms an invaluable asset.
Unit Economics and Cost Control
In the era of Generative AI, product-market fit no longer automatically guarantees a durable business. Operating AI infrastructure is expensive, leading to a risk of “scaling to bankruptcy.” Qiao highlights that customized models can be 5 to 10 times cheaper per task compared to black-box APIs, making them essential for achieving positive gross margins and sustainable growth, especially for digital natives with large user bases.
Security and the Open vs. Closed Debate
Qiao views open models as crucial for striking a balance in cybersecurity by encouraging a broader community to contribute to defense mechanisms. She cites an incident where an open model (GLM2) helped resolve a cyber security breach that a closed model (OpenAI’s) failed to address.
She dismisses concerns about American companies using Chinese open-source models, advocating strongly for the principle of open intelligence regardless of origin, much like the open-source data projects that historically propelled the data science field. Qiao questions why American companies, particularly market leaders like OpenAI and Anthropic, should not open-source their models, believing they should set an example for the industry.
Fireworks’ Moat and Operational Excellence
Fireworks positions itself as a specialized intelligence platform, not merely an inference provider or a cloud service. Its long-term moat lies in its co-optimized training and inference platform, designed to maximize quality, speed, and cost for customized solutions. Fireworks is known for its “day zero launches” of new models, sometimes even reverse-engineering them, but always prioritizing quality; for instance, delaying a DeepSeek 4 launch by three days to fix bugs and contribute fixes back to the open-source community. This rapid deployment enables customers to quickly validate new base models against their internal evaluations.
Fireworks’ proprietary training and inference engines are designed for extreme modularity and optimization (offering over 100,000 inference options), allowing it to deliver “bitwise equivalent” precision between training and inference. While not open-sourcing its core engine due to the rapid pace of development and the complexity of managing external contributions, Fireworks actively collaborates and provides feedback to open-source communities and hardware vendors.
Leadership and Authenticity
Lin Qiao describes her journey as an entrepreneur, driven by a desire for industry-wide impact. She credits her experience at Facebook for instilling a deep sense of extreme ownership. As a founder, she values authenticity, transparency, and intellectual honesty within her team, frequently conducting “pre-mortems” to anticipate challenges. Her biggest surprise as a CEO has been the lack of sufficient data for early-stage decision-making, necessitating reliance on gut calls and intuition, with a strong emphasis on continuous validation and adapting quickly. She champions an authentic marketing approach, trusting the company’s product and customer relationships to speak for themselves.